Spectral Bias in Practice: The Role of Function Frequency in Generalization

Authors: Sara Fridovich-Keil, Raphael Gontijo Lopes, Rebecca Roelofs

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we propose methodologies for measuring spectral bias in modern image classification networks on CIFAR-10 and Image Net. We find that these networks indeed exhibit spectral bias, and that interventions that improve test accuracy on CIFAR-10 tend to produce learned functions that have higher frequencies overall but lower frequencies in the vicinity of examples from each class.
Researcher Affiliation Collaboration Sara Fridovich-Keil University of California, Berkeley sfk@eecs.berkeley.edu Raphael Gontijo-Lopes Google Brain iraphael@google.com Rebecca Roelofs Google Brain rofls@google.com
Pseudocode No The paper describes the methods in narrative text and mathematical formulations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Our CIFAR-10 code is available at https://github.com/google-research/google-research/tree/master/spectral_bias; for Image Net we apply the same interpolation method to pre-cropped images and pretrained model checkpoints.
Open Datasets Yes We use the CIFAR-10 [23] dataset of low-resolution (32 32) natural images from ten animal and object classes and the Image Net [8] dataset of higher-resolution (224 224) natural images from 1000 classes.
Dataset Splits Yes Our label smoothing experiments use the CIFAR-10 dataset [23], where ntrain = 50000, nval = 10000, d = 32, c = 3, and M = 10.
Hardware Specification No These results did require substantial GPU compute, but we do not quantify the exact amount. Each point in a label smoothing figure required training a model from scratch, and each value in an interpolation figure required evaluating a pretrained model on hundreds of thousands of images.
Software Dependencies No The paper does not explicitly list software dependencies with version numbers, such as Python or specific deep learning frameworks and libraries.
Experiment Setup Yes We train from scratch using the original examples Xi and their smoothed labels yi. We train a Wide Res Net32 model (wide-resnet with width 32) with radial wave label smoothing at frequency 0.04. ... training we tested. ... Mixup (with strength 0.1) ... weight decay and early stopping.